“The UK's AI Safety Institute has successfully identified vulnerabilities in AI systems that the private labs which created them had failed to find.”